ggml-cpu : fix leftover handling in ggml_vec_scale_f32 for SVE (llama/16443)

This commit updates the leftover handling in ggml_vec_scale_f32.

The motivation for this is that the code currently incorrectly assumes
there would be fewer than ggml_f32_epr leftover elements. However,
since the main loop processes 2*ggml_f32_epr elements per iteration
, there can be up to (2*ggml_f32_epr - 1) leftover elements.

The original single-pass leftover code could only process ggml_f32_epr
elements, leaving some elements unscaled.

Example scenario with 256-bit SVE:
```
ggml_f32_epr  = 8 (elements per register)
ggml_f32_step = 16 (two registers per iteration)
n             = 25
np            = 16
leftovers     = 9 elements (16-24)

Original    : processes only elements 16-23, misses element 24
This commit : loop processes elements 16-23, then element 24
```

Refs: https://github.com/ggml-org/llama.cpp/actions/runs/18070620247/job/51419855630

diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h
index 341e64e64f2b0da08b96cc078ec805603bd19607..f95ca94e54b16cbf27b64f1d0ba023cfd78da913 100644 (file)
--- a/ggml/src/ggml-cpu/vec.h
+++ b/ggml/src/ggml-cpu/vec.h
@@ -654,11 +654,11 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
         }
         // leftovers
         // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
-        if (np < n) {
-            svbool_t pg = svwhilelt_b32(np, n);
-            ay1 = svld1_f32(pg, y + np);
+        for (int i = np; i < n; i += ggml_f32_epr) {
+            svbool_t pg = svwhilelt_b32(i, n);
+            ay1 = svld1_f32(pg, y + i);
             ay1 = svmul_f32_m(pg, ay1, vx);
-            svst1_f32(pg, y + np, ay1);
+            svst1_f32(pg, y + i, ay1);
         }
     #elif defined(__riscv_v_intrinsic)
         for (int i = 0, avl; i < n; i += avl) {